1 / 44

Performance and Extension of User Space File

Performance and Extension of User Space File. Aditya Raigarhia and Ashish Gehani Stanford and SRI. ACM Symposium on Applied Computing (SAC) Sierre , Switzerland, March 22-26, 2010. Introduction (1 of 2). Developing in-kernel file systems challenging

alijah
Download Presentation

Performance and Extension of User Space File

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Performance and Extension of User Space File Aditya Raigarhia and Ashish Gehani Stanford and SRI ACM Symposium on Applied Computing (SAC) Sierre, Switzerland, March 22-26, 2010

  2. Introduction (1 of 2) • Developing in-kernel file systems challenging • Understand and deal with kernel code and data structures • Steep learning curve for kernel development • No memory protection • No use of debuggers • Must be in C • No standard C library • In-kernel implementations not so great • Porting to other flavors of Unix can be difficult • Needs root to mount – tough to use/test on servers

  3. Introduction (2 of 2) • Modern file system research adds functionality over basic systems, rather than designing low-level systems • Ceph [37] – distributed file system for performance and reliability – uses client in users space • Programming in user space advantages • Wide range of languages • Use of 3rd party tools/libraries • Fewer kernel quirks (although still need to couple user code to kernel system calls)

  4. Introduction - FUSE • File system in USEr space (FUSE) – framework for Unix-like OSes • Allows non-root users to develop file systems in user space • API for interface with kernel, using fs-type operations • Many different programming language bindings • FUSE file systems can be mounted by non-root users • Can compile without re-compiling kernel • Examples • WikipediaFS [2] lets users view/edit Wikipedia articles as if local files • SSHFS access via SFTP protocol

  5. Problem Statement • Prevailing view – user space file systems suffer significantly lower performance compared to kernel • Overhead from context switch, memory copies • Perhaps changed due to processor, memory and bus speeds? • Regular enhancements also contribute to performance? • Either way – measurement of “prevailing view”

  6. Outline • Introduction (done) • Background (next) • FUSE overview • Programming for FS • Benchmarking • Results • Conclusion

  7. Background – Operating Systems • Microkernel (Mach [10], Spring [11]) have only basic services in kernel • File systems (and other services) in user space • But performance is an issue, not widely deployed • Extensible OSes (Spin[1], Vino[4]) export OS interfaces • User level code can modify run-time behavior • Still in research phase

  8. Background – Stackable FS • Stackable file systems [28] allow new features to be added incrementally • FiST [40] allows file systems to be described using high-level language • Code generation makes kernel modules – no recompilation required • But • cannot do low-level operations (e.g., block layout on disk, metatdata for i-nodes) • Still require root to load

  9. Background – NFS Loopback • NFS loopback servers [24] puts server in user-space with client • Provides portability • Good performance • But • Limited to NFS weak cache consistency • Uses OS network stack, which can limit performance

  10. Background - Misc • Coda [29] is distributes file system • Venus cache manager in user space • Arla [38] has AFS user-space daemon • But not widespread • ptrace() – process trace • Working infrastructure for user-level FS • Can interacept anything • But significant overhead • puffs [15] similar to FUSE but NetBSD • FUSE built on puffs for some systems • But puffs not as widespreadh

  11. Background – FUSE contrast • FUSE similar since loadable kernel module • Unlike others is mainstream – part of Linux since 2.6.14, ports to Mac OSX, OpenSolaris, FreeBSD and NetBSD • Reduces risk of obsolete once developed • Licensing flexible – free and commercial • Widely used (examples next)

  12. Background – FUSE in Use • TierStore [6] distributed file system to simply deployment of apps in unreliable networks • Uses FUSE • Increasing trend for dual OS (Win/Linux) • NTFS-3G [25] open source NTFS uses FUSE • ZFS-FUSE [41] is port of Zeta FS to Linux • VMWare disk mount [36] uses FUSE on Linux

  13. FUSE Example – SSHFS on Linux https://help.ubuntu.com/community/SSHFS % mkdir ccc % sshfs -o idmap=user claypool@ccc.wpi.edu:/home/claypool ccc % fusermount -u ccc

  14. Outline • Introduction (done) • Background (done) • FUSE overview (next) • Programming for FS • Benchmarking • Results • Conclusion

  15. FUSE Overview • On userfsmount, FUSE kernel module registers with VFS • e.g., call to “sshfs” • userfsprovides callback functions • All file system calls (e.g., read()) proceed normally from other process • When targeted at FUSE dir, go through FUSE module • If in page cache, return • Otherwise, to userfsvia /dev/fuse and libfuse • userfscan do anything (e.g., request data from ext3 and add stuff) before returning data • fusermount allows non-root users to mount

  16. FUSE APIs for User FS • Low-level • Resembles VFS – user fs handles i-nodes, pathname translations, fill buffer, etc. • Useful for “from scratch” file systems (e.g., ZFS-FUSE) • High-level • Resembles system calls • User fs only deals with pathnames, not i-nodes • libfuse does i-node to path translation, fill buffer • Useful when adding additional functionality

  17. FUSE – Hello World Example ~/fuse/example$ mkdir /tmp/fuse~/fuse/example$ ./hello /tmp/fuse~/fuse/example$ ls -l /tmp/fusetotal 0-r--r--r-- 1 root root 13 Jan 1 1970 hello~/fuse/example$ cat /tmp/fuse/helloHello World!~/fuse/example$ fusermount -u /tmp/fuse~/fuse/example$ Flow Run http://fuse.sourceforge.net/helloworld.html

  18. FUSE – Hello World (1 of 4) Callback operations Invoking does ‘mount’ http://fuse.sourceforge.net/helloworld.html

  19. FUSE – Hello World (2 of 4) Fill in file status structure (type, permissions) http://fuse.sourceforge.net/helloworld.html

  20. Check that path is right Check permissions right (read only) Check that path is right Copy data to buffer http://fuse.sourceforge.net/helloworld.html

  21. FUSE – Hello World (4 of 4) Copy in directory listings http://fuse.sourceforge.net/helloworld.html

  22. Performance Overhead of FUSE : Switching • When using native (e.g., ext3) • Two user-kernel mode switches (to and from) • Relatively fast since only privilege/unpriviledge • No context switches between processes/address space • When using FUSE • Four user-kernel mode switches (adds up to userfs and back) • Two context switches (user process and userfs) • Cost depends upon cores, registers, page table, pipeline

  23. Performance Overhead of FUSE : Reading • FUSE used to have 4 KB read size • If memory constrained, large reads would do many context switch each read • swap out userfs, bring in page, swap in userfs, continue request, swap out userfs, bring in next page … • FUSE now reads in 128 KB chunks with big_writes mount option • Most Unix utilities (cp, cat, tar) use 32 KB file buffers

  24. Performance Overhead of FUSE : Time for Writing (Write 16 MB file) Note, benefit from 4KB to 32KB, but not 32KB to 128KB

  25. Performance Overhead of FUSE : Memory Copying • For native (e.g., ext3), write copies from application to kernel page cache (1x) • For user fs, write copies from application to page cache, then from page cache to libfuse, then libfuseto userfs(3x) • direct_io mount option – bypass page cache, user copy directly to userfs(1x) • But reads can never come from kernel page cache!

  26. Performance Overhead of FUSE : Memory Cache • For native (e.g., ext3), read/written data in page cache • For user fs, libfuseand userfsboth have data in page cache, too (extra copies) – useful since make overall more efficient, but reduce size of usable cache

  27. Outline • Introduction (done) • Background (done) • FUSE overview (done) • Programming for FS (next) • Benchmarking • Results • Conclusion

  28. Language Bindings • 20 language bindings – can build userfsin many languages • C++ or C# for high-perf, OO • Haskell and OCaml for higher order functions (functional languages) • Erlangfor fault tolerant, real-time, distributed (parallel programming) • Python for rapid development (many libraries) • JavaFuse [27] built by authors

  29. Java Fuse • Provides Java interface using Java Native Interface (JNI) to communicate from Java to C • Developer writes file system as Java class • Register with JavaFuse using command line parameter • JavaFusegets callback, sends to Java class • Note, C to Java may mean more copies • Could have “file” meta-data only option • Could use JNI non-blocking I/O package to avoid But both limit portability and are not thread safe

  30. Outline • Introduction (done) • Background (done) • FUSE overview (done) • Programming for FS (done) • Benchmarking (next) • Results • Conclusion

  31. Benchmarking Methodology (1 of 2) • Microbenchmarks – raw throughput of low-level operations (e.g., read()) • Use Bonnie [3], basic OS benchmark tool • 6 phases: • write file with putc(), one char at a time • write from scratch same file, with 16 KB blocks • read file with getc(), one char at a time • read file, with 16 KB blocks • clear cache, repeat getc(), one char at a time • clear cache, repeat read with 16 KB blocks

  32. Benchmarking Methodology (2 of 2) • Macrobenchmarks – application performance for common apps • Use Postmark – small file workloads by email, news, Web-based commerce under heavy load • Heavy stress of file and meta-data • Generate initial pool of random text files (used 5000, from 1 KB to 64 KB) • Do transactions on files (used 50k transactions, typical Web server workload [39]) • Large file copy – copy 1.1 GB movie using cp • Single, large file as for typical desktop computer or server

  33. Testbed Configurations • Native – ext4 file system • FUSE – null FUSE file system in C (passes each call to native file system) • JavaFuse1 (metadata-only) – null file system in JavaFuse, does not copy read()and write()over JNI • JavaFuse2 (copy all data) – copies all over JNI

  34. Experimental Setup • P4, 3.4 GHz, 512 MB RAM (increased to 2 GB for macrobenchmarks), 320 GB Seagate HD • Maximum sustained throughput on disk is 115 MB/s • So, any reported throughputs above 115 MB/s must benefit from cache • Linux 2.6.30.5, FUSE 2.8.0-pre1

  35. Outline • Introduction (done) • Background (done) • FUSE overview (done) • Programming for FS (done) • Benchmarking (done) • Results (next) • Conclusion

  36. Microbenchmark Results • FUSE~25% overhead • Context switches • JavaFusemore overhead (context switches between JNI and C), and doing copies not much worse than metadata

  37. Microbenchmark Results • Native much faster than FUSEfor small files (written to memory) • For large cache, written to disk which dominates

  38. Microbenchmark Results • Data already in page cache, so fast for all • Spikes are when 512 MB RAM exhausted (FUSE has two copies, so earlier around 225)

  39. Microbenchmark Results • For native, can cached up to 500 MB • For Java, spike is caused by artificial nature of benchmark (previous version still in cache)

  40. Microbenchmark Results input • When cache cleared, starts lower gets higher for larger • Kernel optimizes for sequential read • FUSE does better – think it’s because fusefsand ext4 both read-ahead

  41. Microbenchmark Results • Native gets same benefit as FUSE, so no apparent difference

  42. Postmark Macrobenchmark Results • FUSE overhead less than 10% • Java overhead about 60% • Greater CPU and memory use

  43. Copy 1.1 GB File Macrobenchmark Results • FUSE comparable to native (~30%) • Java overhead minimal over FUSE

  44. Conclusion • FUSE may be feasible depending upon workload • Performance comparable to in-kernel for large, sustained I/O • Overhead noticeable for large number of meta data (e.g., Web and many clients) • Adequate for PCs and small servers for I/O transfer • Additional language (Java) incurred additional overhead • But overhead can be reduced with optimizations (e.g., shared buffers)

More Related